AI · Verification
VeriLens.
Production-grade multimodal fact-checking engine that cross-examines claims and images against live reporting, then returns an explainable Real, Fake, or Suspicious verdict with a 0–100 credibility score.
The brief
VeriLens is an AI-driven fact-checking engine that treats verification as active cross-examination rather than black-box classification. It retrieves contemporaneous reporting, checks contextual consistency between text claims and accompanying imagery, and generates transparent rationales backed by primary sources. The TypeScript and Express backend combines GPT-4o-mini, Gemini 2.5 Flash, SerpAPI news retrieval, Cloudinary media handling, and deterministic scoring. JWT authentication, refresh-token rotation, and role-based access protect user histories and administrative controls.
Capabilities
- 01Multimodal Claim Verification
- Accepts claim text and optional JPG, PNG, or WebP evidence, then returns a Real, Fake, or Suspicious verdict with a 0–100 credibility score and an explainable journalistic breakdown.
- 02Live Journalism Cross-Referencing
- Extracts high-signal keywords, queries SerpAPI Google News, and evaluates publisher identity, timestamps, consensus, and contradictions across contemporaneous reporting.
- 03Multi-LLM Fallback Architecture
- Routes requests first to GPT-4o-mini and falls back to Gemini 2.5 Flash during rate limits or network failures. Strict JSON-schema validation prevents AI formatting drift.
- 04Forensic Image Alignment
- Streams uploads through Cloudinary and assesses whether imagery genuinely depicts the claimed event, including temporal inconsistencies and synthetic-generation artifacts.
- 05Deterministic Scoring
- Combines text confidence and visual alignment with a 60/40 weighted formula. Scores below 30 classify as Fake, 30–64 as Suspicious, and 65 or higher as Real, with overrides for verified debunks.
- 06Enterprise Security
- Uses short-lived access tokens, seven-day refresh-token rotation, Bcrypt hashing, email verification, password reset, and role-based access control.
- 07History and Administration
- Persists queries, model outputs, sources, breakdowns, confidence scores, and timestamps. Role-checked administrative routes support account auditing and permission promotion.